An Enhanced Classification Approach using Hyperspectral Image Data in Combination with in situ Spectral Measurements for the Mapping of Vegetation Communities

Abstract

This paper shows the potential of a method using field spectral measurements as independent training data for the classification of airborne hyperspectral imagery of a natural preserve in Germany, using two different machine learning algorithms. The spectral reflectance of different dry grass- and heathland vegetation communities was measured with field spectrometers (350 nm - 2500 nm) in August 2009. Additionally, hyperspectral imagery was acquired by the airborne scanner aisaEAGLE (390 nm - 970 nm). The developed normalization technique was proven to be a suitable method to make image and field spectra comparable for classification. A support vector machine (SVM) and random forest (RF) classifier trained with normalized field spectra were applied to normalized image data to classify thy grass- and heathland communities in different levels of detail. SVM (overall accuracy (OAA) 89.13%) provided significantly better classification results compared to RF (OAA 71.74%) in the second level of detail. Consequently, only SVM was used for classification in the highest level of detail (third level), which also led to high classification accuracy (OAA 77.27%). The results indicate the potential of the developed approach classifying airborne hyperspectral image data with field spectral measurements for the spatial assessment and separation of dry grass- and heathland communities.